GS Caltex cuts model deployment time from a week to hours using SageMaker

South Korean energy and chemical company GS Caltex has streamlined its machine learning operations by connecting no-code tools to production workflows. The firm divides its machine learning work between process engineers, who understand the physical chemical operations, and machine learning engineers, who manage infrastructure. By automating the handoff between these two groups, GS Caltex reduced the time required to deploy new machine learning models from roughly a week to a few hours.

The company also changed how it generates predictions. Instead of running continuous server endpoints, GS Caltex uses scheduled batch jobs every 30 minutes. This approach has meaningfully lowered its ongoing inference costs while still serving the needs of its operational teams.

Bridging no-code tools and production workflows

Process engineers at GS Caltex build and validate machine learning models using Amazon SageMaker Canvas. This visual interface allows them to apply their specialized knowledge to prepare data and evaluate predictions without writing code. They use an integrated data tool called Data Wrangler to define transformation steps, such as cleanups and scalings, and evaluate whether a model’s outputs are useful for real-world operations.

However, moving these assets into a reliable production pipeline traditionally required significant manual integration. Machine learning engineers had to manually update data paths, match container images, configure models, and script batch processing tasks. GS Caltex created an automated handoff system using Amazon SageMaker Pipelines to convert the visual assets directly into repeatable, executable production steps.

Standardizing the handoff process

The automated production system relies on two distinct phases: preparing a new model version and running scheduled batch inference. When a process engineer finishes building a model, the system registers it in a central repository called the Amazon SageMaker Model Registry. The underlying data preparation logic is saved as a specific flow file in Amazon Simple Storage Service (Amazon S3).

Machine learning engineers review the data formats, check validation results, and formally approve the model inside the registry. Once approved, an automated workflow takes over. The system reuses the exact preprocessing steps validated by the process engineer, ensuring that scaling and encoding parameters learned during training are preserved rather than calculated again on new data.

Automating batch inference to reduce overhead

To run predictions without paying for constantly running cloud servers, GS Caltex uses Amazon EventBridge to trigger batch runs every 30 minutes. When triggered, SageMaker Pipelines retrieves raw input data from Amazon S3, executes the saved preprocessing flow through SageMaker Processing, and runs predictions using SageMaker Batch Transform.

• Process engineers create models visually without writing code.
• Machine learning engineers approve assets and maintain production schedules.
• Scheduled triggers execute preprocessing and inference automatically every half hour.
• Output files are stored back in Amazon S3 for downstream use.

Machine learning engineers monitor the whole operation by inspecting pipeline step statuses and system logs in Amazon CloudWatch. If a failure occurs, the operational metrics make it easy to spot whether the issue stem from data formatting, S3 permissions, or container deployment.

Why it matters for builders and investors

For technology builders, the GS Caltex architecture offers a practical pattern for scaling artificial intelligence across domain experts. Companies often struggle with a bottleneck where specialized machine learning engineers must build every data pipeline from scratch. By enabling domain experts to build validated models visually while automatically capturing their data transformation flows, engineering teams can focus on platform reliability rather than basic model building.

For investors, this implementation highlights a shift toward cost-efficient AI operations. Continuous real-time endpoints can quickly become expensive, especially for industrial use cases that only require periodic updates. Standardizing on scheduled batch transform jobs cuts compute infrastructure costs while maintaining operational performance.

What to watch next

As enterprise adoption of no-code AI tools grows, monitor how cloud vendors automate the bridge between developer environments and operational pipelines. Key areas to track include seamless region-to-region deployment tools, automated validation checks for incoming data, and tighter management of feature parameters across training and batch production environments.